Cost of Goods Sold in AI SaaS: The Number Vendors Do Not Report

Gross margin decides how an AI company is valued, and in AI it is not really a measurement. It is a classification decision. C3.ai closed fiscal 2026 with a 31% GAAP gross margin and told the market about a 46% non-GAAP one. Inference, retrieval, evaluation and human review all sit inside that same caption, and no accounting rule says exactly where each of them has to land. AI COGS is where the discretion lives.

The margin you are shown and the margin that exists

Ask a buyer what a vendor's gross margin tells them and you get a reasonable answer. It tells you whether the business works. It tells you how much room there is on price. It tells you whether the vendor can survive a bad year without repricing you.

All of that is true when the number means the same thing at every company. In AI it does not.

Two vendors with identical operations can report margins ten to fifteen points apart, legally, audited, with no disagreement between them about a single dollar of spend. They disagree only about which caption each dollar belongs in. That is not a loophole. It is what the reporting rules actually ask for.

My view is that gross margin has quietly become the least comparable headline metric in enterprise software. It still gets used as though it were the most comparable one.

What actually sits inside AI cost of revenue

Start with a vendor that writes it down. C3.ai defines the caption in its own filings. Cost of subscription revenue consists primarily of compensation for production environment, support and centre of excellence staff, hosting of its software including payments to outside cloud service providers, and allocated overhead and depreciation for facilities.

Read that list again. Four of the items are people and buildings. One of them is the cloud bill. The compute everybody talks about is a single element inside a caption that is mostly payroll.

The five lines that were not there before

Classic software carried hosting and support in cost of revenue and very little else. AI products added five cost lines that a 2015 SaaS income statement simply did not have.

The five additions are model inference compute, retrieval and vector infrastructure, evaluation and monitoring pipelines, human review of model output, and the ongoing data engineering that keeps retrieval accurate.

Each is a genuine delivery cost. Without any of them the customer does not get a working product. That is the test cost accounting is supposed to apply.

Why classification is a choice and not a rule

Regulation S-X Rule 5-03 sets out how a registrant presents its income statement. It requires expenses to be classified by function, meaning cost of revenue, selling, general and administrative, and research and development. It does not tell you which specific activities are which.

That distinction carries the whole argument. Inference used to serve a paying customer is a delivery cost. Inference used to evaluate a candidate model is research. The same GPU, the same hour, the same invoice, two captions.

An auditor tests whether the policy is reasonable and consistently applied. It does not re-run the allocation. So a vendor that classifies its evaluation and safety compute as research and development is not doing anything improper, and its gross margin is several points higher than an otherwise identical vendor that does not.

A worked example: is the margin 31% or 46%

C3.ai is useful here because it publishes both numbers on the same page and shows its work.

For the fiscal year ended 30 April 2026, the company reported total revenue of $250.3 million and GAAP gross profit of $77.4 million. That is a 31% GAAP gross margin. In the same release it presented non-GAAP gross profit of $116.2 million, a 46% margin.

The 15 point difference is not a rounding artefact or a one-off charge. It is $37.5 million of stock based compensation and $1.3 million of employer payroll tax on that compensation, both sitting inside cost of revenue and both removed for the adjusted figure.

Those are real people doing real delivery work. The company pays them in stock instead of cash. The cost of serving customers did not fall because of how the payroll was funded.

Cost that does not fall when revenue falls

The more instructive number is the direction of travel.

Subscription revenue fell from $327.6 million to $227.1 million. Cost of subscription revenue went the other way, from $143.8 million to $166.3 million. Subscription gross margin therefore dropped from about 56% to about 27% in a single year.

Classic SaaS does not behave like this. Losing a third of your revenue in traditional software costs you a third of your hosting bill and very little else. In an AI product the delivery base is committed capacity and specialist headcount, and neither of those unwinds in twelve months.

That asymmetry is the part buyers should care about. A vendor whose costs do not fall with revenue has a strong incentive to reprice its remaining customers, and you are one of them. The mechanics of that repricing are covered in the analysis of seat compression and AI pricing.

What the accounting rules require, and what they leave open

There are two rules worth knowing, and neither does what buyers assume.

The first is Rule 5-03 of Regulation S-X, already described. It fixes the shape of the income statement and requires separate presentation of product and service revenue when either exceeds 10% of net sales. It says nothing about which costs constitute delivery.

The second is the SEC's non-GAAP framework. An adjusted gross margin is a non-GAAP measure and must be reconciled to the GAAP figure, presented with no more prominence than it, and not described as though it were GAAP. C3.ai complies with all of that. Compliance is not the issue. The issue is that the adjusted number is the one that ends up in the headline and the deck.

What the new disaggregation rule will show, and what it will still hide

In November 2024 the FASB issued ASU 2024-03. It requires public companies to break relevant income statement captions into five named categories. Cost of revenue is one of the captions in scope.

The five categories are purchases of inventory, employee compensation, depreciation, amortisation of intangibles, and depletion. It takes effect for annual periods beginning after 15 December 2026, with interim reporting a year later.

This is a genuine improvement. For the first time you will see how much of a software company's cost of revenue is payroll. In C3.ai's case that would confirm what the policy note already implies, which is that the caption is mostly people.

Here is what it will not do. Purchased inference is a service bought from a cloud provider. It is not inventory, not employee compensation, not depreciation and not amortisation. The single largest question about AI cost of revenue is not one of the five things the new rule makes you name.

I would rather the standard had included purchased services as a sixth category. It did not, and the practical consequence is that the disclosure most relevant to AI economics arrives as a residual rather than a line item.

The private company problem: nobody has published an audited number

Almost every AI gross margin figure in circulation is an estimate produced by somebody outside the company.

Reporting in June 2026 traced the widely quoted Anthropic figures to PitchBook analysis of computing spend per dollar of revenue, at $0.71 in the first quarter of 2026. The same reporting made the point that matters more than the number: an Anthropic S-1 would carry the first audited gross margin ever published by a frontier AI lab.

Until that document exists, every margin figure for a private AI company is a construction. Different analysts include different things, which is exactly why published estimates for the same company in the same year differ by twenty points or more.

This is not a reason to ignore the estimates. It is a reason to read whose estimate it is and what went into the denominator before quoting it in a board paper. The gap between reported and audited numbers is a recurring theme in the review of which AI revenue figures are actually verifiable.

Survey data has the same problem in a friendlier form. ICONIQ's 2026 work put average gross margin across its sample at 45% for 2025, with higher figures projected for later years. The sample is roughly 300 executives, partly drawn from its own portfolio, and the margins are self reported against no common definition. That is useful directional evidence and it is not an audited benchmark.

How to read a vendor's real margin structure

You cannot recalculate a vendor's gross margin from outside. You can work out how much confidence the stated number deserves, and that is usually enough.

Four adjustments to make yourself

For a public vendor, four moves get you closer to a comparable figure in about twenty minutes.

First, use the GAAP margin as your baseline and treat the adjusted one as marketing. Second, add back the stock compensation that sits in cost of revenue, since the reconciliation table tells you the amount and the work it paid for still happened.

Third, read the accounting policy note for cost of revenue and write down what is missing from the list. Human review, solutions engineering and customer specific tuning are the usual absences. Fourth, check the purchase commitments note for cloud capacity, because a large multi-year commitment is a cost that is already locked whatever happens to demand.

For a private vendor none of this is available, and the honest position is that you are relying on what they choose to tell you. In that case the questions in the table above are the whole toolkit. How a vendor answers them is itself informative, and this is one of the few places where the quality of the answer matters more than its content.

Where this argument is weakest

There is a serious counter-case, and it is worth stating properly rather than as a token caveat.

Bain Capital Ventures argued in 2024 that a low AI gross margin today says very little about the margin at scale. Their reasoning is specific. Inference prices fall sharply and unpredictably. They cite a case where a model's price dropped from $2 to $0.24 per million tokens within days. On their arithmetic that moves a 60% margin to 78% with no operational change at all.

They also make a point I find genuinely persuasive. Some AI companies put human labour into COGS deliberately, because the review work generates proprietary data that becomes a defensible asset. On that reading, a deliberately depressed gross margin is an investment decision, not a weakness. Compressing it too early would be the actual error.

The second weakness is my main example. C3.ai's fiscal 2026 was a year of contracting revenue and business model transition, so its 31% is not a proxy for the category. It is a clean illustration of the mechanism because the company disclosed the components. It is not a benchmark.

Third, none of this is unique to AI. Software companies have classified customer success into three different captions for two decades. What AI changed is the size of the discretionary amount, not the existence of the discretion.

The fair summary is that gross margin in AI is less comparable and less predictive than it used to be. That does not make it useless. It makes it a starting question rather than a finishing one, and reading it as a finishing one is the actual error I see buyers make.

Frequently asked questions

Before the next renewal

Pick your two largest AI vendors and spend an hour on each.

If they are public, open the most recent annual filing and find the accounting policy paragraph defining cost of revenue. Copy the list into your notes. Then find the GAAP to non-GAAP reconciliation and write down how much of the gap sits above the gross profit line. You now know the two things that matter: what they count as delivery, and how much they prefer you did not count.

If they are private, send the six questions in the table. Send them in writing, before the commercial conversation starts, and treat the response time as data. A vendor that answers in two days has the numbers. A vendor that routes it to a sales engineer for a fortnight does not.

Either way you are asking about their cost structure rather than your discount, which is where the durable part of the negotiation lives. The same discipline applied to compute economics is set out in the piece on why falling token prices have not produced falling bills. The accounting side of the same question sits in the analysis of GPU depreciation schedules.

Related reading